TL;DR
This paper reviews the current landscape of algorithms for estimating breathing rate (BR) from electrocardiogram (ECG) and photoplethysmogram (PPG) signals. It provides a structured overview of these algorithms, proposes a framework for their assessment, and highlights crucial future research directions for their clinical and wearable applications.
What Problem Does This Paper Address?
Breathing rate (BR) is a vital physiological indicator of health, yet it's often manually counted, a process that is labor-intensive, prone to error, and unsuitable for continuous or unobtrusive monitoring. The paper addresses the need for automated, electronic, and unobtrusive methods for BR estimation, particularly by leveraging readily available ECG and PPG signals, which are commonly captured in healthcare and fitness devices. This would enable earlier detection of health deteriorations and improve patient monitoring.
Methods
The paper categorizes BR algorithms into a five-stage structure:
- Extract Respiratory Signal(s): This involves isolating signals dominated by respiratory modulations from the primary cardiac signals (ECG/PPG). Techniques include filter-based methods (e.g., bandpass filtering) and feature-based methods (e.g., extracting beat-by-beat amplitude variations).
- Fuse Respiratory Signals (Optional): Multiple extracted respiratory signals can be combined to create a single, more robust signal.
- Estimate BR(s): BR is estimated from a window of the respiratory signal using time-domain (e.g., breath detection) or frequency-domain (e.g., spectral analysis) techniques.
- Fuse BR(s) (Optional): Multiple BR estimates can be combined to improve the final estimate's robustness.
- Quality Assessment (Optional): Signal quality indices (SQIs) and respiratory quality indices (RQIs) are used to identify and potentially reject imprecise estimates.
The paper also reviews experimental methodologies for assessing BR algorithms and proposes a methodological framework for future assessments.
Datasets
The paper discusses the use of various datasets for evaluating BR algorithms. It notes that many studies use datasets that are not representative of target populations and are often not publicly available. A total of 13 publicly available datasets were identified as having been used to assess BR algorithms. The paper lists these datasets in Table VI, detailing their characteristics such as age range, signal types used (ECG, PPG, etc.), type of breathing, and level of illness of subjects.
Key Insights and Ideas
- Ubiquitous Signals, Untapped Potential: ECG and PPG, already common in wearables and clinical settings, offer a significant opportunity for unobtrusive BR monitoring without additional sensors.
- Respiratory Modulations are Key: The subtle ways respiration affects ECG (baseline wander, amplitude modulation, frequency modulation) and PPG are the basis for most BR estimation algorithms.
- No One-Size-Fits-All: Different patient populations and clinical settings may require different BR algorithms due to variations in respiratory physiology and desired performance metrics (e.g., precision vs. availability of estimates).
- The Challenge of Validation: Comparing algorithm performance across studies is difficult due to varied statistical measures, subject populations, and lack of standardized implementations.
- Quality Matters: Beyond signal quality, assessing the respiratory quality of signals (RQIs) is crucial, as respiratory modulations are independent of overall signal quality and can be affected by various physiological factors.
Key Results
- A comprehensive review identified 196 publications on BR algorithms, with a significant increase in research since 1999.
- Most algorithms use either ECG (50.0%) or PPG (57.1%) as input, with fewer utilizing both or pulse transit time.
- Window durations commonly used range from 30 to 90 seconds, though a trend towards lower errors with longer durations was observed.
- Datasets often comprise young, healthy subjects, with fewer studies using elderly, chronically ill, or acutely ill patient populations.
- The paper highlights the difficulty in determining the "best" performing algorithms due to inconsistent evaluation methods.
Limitations and Future Work
- Limitations: The paper points out the lack of standardized implementations and the difficulty in comparing algorithm performance across studies due to varied statistical measures and subject populations. Many datasets are not representative of target populations and are not publicly available.
- Future Work:
- Developing BR algorithms that incorporate models of respiratory modulations.
- Exploring fusion techniques for combining BR estimates from multiple algorithms.
- Leveraging machine learning techniques as more annotated data becomes available.
- Quantifying the uncertainty associated with BR estimates to discard unreliable ones.
- Developing low-computational requirement algorithms for wearable sensors.
- Investigating BR estimation from video-based PPG.
- Conducting clinical trials to determine the benefit of BR algorithms in various patient populations and settings.
- Assessing the impact of cardiac arrhythmias on algorithm performance.
- Exploring the use of BR algorithms for estimating BR variability as a marker of mental state and disease progression.
Summary updated